Starting release tests.

Caused problems on Mac OS X 10.6.8.
This fixes the forgotten bits in commit
fa046011

Now possible to use work group size 1. Should have been part of
commit 745e3400 already.

Resolves #149.
Works for dense matrices. Sparse matrices are intentionally not part
of this commit, because sparse matrix transposition is something one
better avoids whenever possible.

Now using two sets of SpMV kernels:
 - One for y = Ax for performance reasons
 - One for y = alpha * Ax + beta * y; for inplace-operations.
Previous use of temporaries for inplace-operations was too slow and
resulted in unnecessary memory overhead.

Required for some OpenCL SDKs on CPUs.

This resulted in errors on Mac OS X 10.6.8.

Work group size 1 is encountered on some older implementations for CPUs
as well as on current mobile hardware.

On CPUs the maximium workgroup size might be 1, which was not considered
in the previous implementation. This commit removes this restriction.

This error is usually encountered if the OpenCL SDK or driver is not
installed correctly. However, this was not properly explained in
the error message. Now fixed.

Since OpenCL is now disabled, viennacl-info is no longer built.

New Ubuntu Trusty machines no longer provide the OpenCL driver from fglrx.
Although this is not perfect, running tests without OpenCL is better
than not running any tests at all...

Apparently the test machines experienced an upgrade, so the old
fglrx version was no longer available.

See http://www.kitware.com/blog/home/post/510 for background info.

- Added new AMD GPU codenames.
- NVIDIA Fermi GPUs are mapped to names in the database.
  Improved detection of Tesla GPUs.

Resolves #150.

karlrupp/feature-reduce-generator-usage:
  Removes BLAS level 1 and 2 operations from generator, only GEMM remains.
  Reasons: no support for char and short, some unnecessary kernel launches, poor CPU support.
  Fixes #167.

Conflicts:
	viennacl/device_specific/builtin_database/devices/gpu/nvidia/kepler/tesla_k20m.hpp

With BLAS level 1 and 2 operations moved out of the generator again,
a lot of code was no longer needed. This commit removes all these
no longer needed code lines.

Operations are now pulled out of the generator, so dead code is no longer needed.

First issue: Internal assertion in Intel OpenCL vectorizer.
             Workaround by refactoring binary element-wise operations.
Second issue: Multiple inner products produced wrong results with CPUs.
              Fixed by using correct number of groups.

Optimization of matrix and vector operations in the OpenMP backend

Squared norm was used in computation, so we need to square the tolerance.

Resolves #166. In short, the OpenCL standard prohibits the use
of floating point arguments for expr1 in the ternary operator
 expr1 ? expr2 : expr3

The boundary for the iteration on coulumns was changed from A_size1 to A_size2

Increases perfomance on my latop due to better memory bandwidth.
Correct a bug if data_alpha is Integer and reciprocal_alpha=TRUE.

Increases performance by 2x due to better memory access.

The matrix will be divided into sub-matrices for better storage access.
This increases the performance by 3x on my laptop.

OpenMp does not support vcl_size_t in for-loops, migration to
long was required.

Increases performance by 2x on my laptop due to better memory bandwidth

Increases performance by 2x on my laptop due to better memory bandwidth
Correct a mistake in the description

Increases performance by 2x on my laptop due to better memory bandwidth.

Increases performance by 2x on my laptop due to better memory bandwidth.
Correct a mistake in the description.

If A has an empty row and less than 4 rows in the row block, then the result
entry corresponding to the empty row is computed incorrectly
(value of next nonzero row).

Reported-by: Patrick Sanan (via work on PETSc)